IterSynth splits deep-search agents into two roles

IterSynth splits deep-search agents into a Planner and a Synthesizer. The two share one evolving summary, and that summary, not the growing search history, is the only state the agent keeps.

That's a problem every long multi-step search agent hits. Tool results pile up until they crowd out the model's own reasoning. IterSynth-8B averages 50.7 across five long-horizon benchmarks including BrowseComp, 4.2% above the strongest prior agent at 8B or smaller, according to the IterSynth paper.

IterSynth-8B is trained with a method the authors call Role-Decoupled Policy Optimization.

The authors say the same Planner/Synthesizer split works as a prompting pattern on an existing model, with no retraining required. Anyone running a long search agent can test the two-role structure without touching weights.

The paper doesn't say how the split holds up past 8B. It also doesn't say whether the shared summary starts losing detail on tasks longer than the five benchmarks tested. The 4.2% gap is an average across all five, not a single best case.